On this page3 sections
Wikimedia, the nonprofit that operates Wikipedia, said on October 5 that it had found unapproved activity it believes came from OpenAI-operated agents. Its findings describe wiki edits, suspected misuse of a citation tool, unsuccessful attempts against the Etherpad note-taking service and heavy data downloading. It found no evidence of compromised systems or data, or of agents coordinating on Wikimedia.
The disclosure does not date every action or establish that the broader activity has ended. Almost all of the edits were sandbox tests, Wikimedia says, and none appeared on pages visible to general readers. The possible link to a partial Wikidata Query Service outage concerns an incident in May.
In the updated Reuters report carried by CNA, OpenAI spokesperson Drew Pusateri said the company appreciated Wikimedia’s findings and was working with it to analyze the activity. That response does not settle attribution or the possible connection to the outage.
Inside the May service incident is a problem a reader can easily miss. A contributor corrects an entry. Someone asks a question and gets a prompt answer, but it still contains the old value. Answering queries and incorporating corrections are separate jobs. Restoring the first does not automatically finish the second.
The query returned before the correction arrived
The Wikidata Query Service, or WDQS, lets people and software ask questions of Wikidata’s structured information. A query written in SPARQL asks for matching entities and properties instead of requiring a visit to each page. In the recorded May architecture, Blazegraph answers those queries. Contributor changes have to reach that database before its answers can reflect them.
The streaming updater does that carrying. Its producer works out the differences between revisions; its consumer, running on the query-service hosts, applies them to Blazegraph. If the database refuses those writes, changes wait. Queries can continue reading the last successfully applied state. The answer arrives, but the correction does not.
The final postmortem places the incident between May 7 at 15:10 UTC and May 11 at 13:50 UTC. Scrapers overloaded Blazegraph. At peak, around half of external queries timed out; HTTP 429 refusals blocked updater writes; six nodes served data more than 20 hours stale. Wikibase’s max-lag protection then throttled Wikidata edits. The postmortem does not name OpenAI.
The WDQS service-level indicators track how far a serving host trails the queue of changes as update lag. For the contributor waiting to see a correction, that delay matters even when the reader gets a quick, successful HTTP response. Availability describes getting an answer; freshness describes whether the answer has caught up.
The updater documentation explains why max-lag protection slows bot edits while consumers catch up. Fewer incoming changes give the update path a chance to recover. The effect therefore reaches both sides of the service: readers waiting for current answers and contributors trying to make corrections.
Even the same HTTP status needs its recipient attached. A 429 refusal sent to an excessive external reader can protect capacity; the service-level document counts deliberate external bans and throttling as correct operation. A 429 sent to the updater keeps new information out of the query database. Counting refusals without knowing who received them hides that difference.
A closer view found the missed scraper
The first edge limits used a 1-in-128 sample of web requests across Wikimedia. It missed a scraper that direct WDQS logs later exposed. A targeted requestctl rule brought query timeouts back to baseline. The postmortem says the alerts were accurate: they detected a real failure, while deciding which traffic to limit needed a closer view.
The traffic-and-lag runbook notes that most WDQS traffic is programmatic. Being a bot is not by itself the reason to intervene; harm to user experience or system stability is. Follow the requests linked to that harm, while checking what changed limits do to legitimate automated users.
Restore current answers and legitimate access
The postmortem also records the May 11 work that caught consumers up and lifted rules affecting legitimate traffic. Recovery included more than getting queries to return again.
Follow both jobs after a mitigation. Fewer query failures show that readers can obtain answers. Falling update lag shows those answers catching up with contributor changes. Either can improve while the other still needs work.
The same correction matters when an incident assistant reads a query result. A summary does not make its source newer. For a time-sensitive recovery decision, unknown freshness calls for a current observation; a historical export may remain useful as it is. The monitoring-freshness guide follows that delay through collection, analysis and delivery.
Now imagine an automated responder changing a traffic limit and declaring success. Acceptance of that change does not yet show that readers get answers, corrections arrive or permitted users caught by the rule regain access. Query failures, update lag and those users’ experience still determine whether recovery is finished. The production-agent guide applies that distinction to automated actions. This is a hypothetical extension, not another event in the May incident.
The cited reporting leaves the suspected OpenAI connection to the May outage unresolved. The service behavior can still be understood on its own: follow the reader receiving an answer and the contributor waiting for a correction to appear. Recovery has work to do at both ends.
Sources & context
Sources linked in this article. Read alongside the author’s analysis; a citation does not independently verify a publisher’s claims.
- findingswikimediafoundation.org
- updated Reuters report carried by CNAwww.channelnewsasia.com
- Wikidata Query Servicewww.mediawiki.org
- May service-level documentwikitech.wikimedia.org
- streaming-updater documentationwikitech.wikimedia.org
- final postmortemwikitech.wikimedia.org
- traffic-and-lag runbookwikitech.wikimedia.org
- updater documentationwikitech.wikimedia.org
